About AI Agents

This page explains what AI agents are, how they work, and when you should use them in your applications. It also provides a guide to help you decide between using an AI agent and an AI skill.

For practical examples of AI agents in action, see Use Cases. To get started building your first agent, see Create and Configure an AI Agent.

What is an AI agent?

An AI agent is a smart, configurable object that can:

• Understand your goals through a natural language prompt.
• Interact with your data and systems using tools.
• Perform tasks like routing cases, updating records, or starting workflows.

Each AI agent follows your instructions to complete tasks consistently.

Agent types

Every AI agent has a type that determines where and how it runs. You select the agent type when you create the AI agent, and it cannot be changed later.

Type	Purpose	Deployment method
Process agent	Automates routine tasks without interruption as part of a business workflow.	Execute AI Agent smart service in a process model
Chat agent	Engages users through natural conversation as a virtual assistant.	Agent chat component (a!agentChatField) in an interface

Runtime validation enforces type boundaries. A process agent cannot be used in the agent chat component, and a chat agent cannot be executed via the Execute AI Agent smart service.

For detailed reference information about each type, see AI Agent Reference.

How AI agents work

AI agents operate using these key elements:

• Prompt: A written description of the tasks the AI agent should perform. This helps the AI agent understand the goal and how to approach it. When you create a new AI agent, you enter your prompt in the Instructions field.
• Tools: Design objects or external capabilities that perform a single specific task. They can let the AI agent access data, trigger processes, call other AI agents, retrieve information, or interact with external systems through MCP tools. Tools ensure AI agents act within structured, secure boundaries. For details on available tools, see AI Agent Screen Reference. For document tool guidance, see Working with Documents in AI Agents.
• Inputs: Data that the AI agent receives to complete a task. Inputs define what information the AI agent needs at the start of its run.
• Outputs: Data that the AI agent returns after completing a task. Outputs are optional—some agents don't produce an explicit result. Instead, they may create side effects during their run, such as writing to a record type or updating data in a process.

You define the goal and set up tools specific to each AI agent, giving you full control over how each AI agent behaves. How you deploy an AI agent depends on its type:

• Process agents run inside a process model using the Execute AI Agent smart service. When the process agent runs, it interprets your prompt, selects the appropriate tools, and executes each step automatically. When the AI agent identifies multiple independent tasks within a single reasoning step, it can execute up to five tool calls in parallel rather than processing them one at a time. Parallel tool execution reduces latency and token costs for multi-tool tasks like research retrieval and multi-source data gathering. For constraints and error handling details, see AI Agent Reference.
• Chat agents run inside an interface using the agent chat component (a!agentChatField). End users interact with the chat agent through a conversational interface, and the AI agent responds in real time.

You can manually stop a running AI agent at any time. This is useful when testing or when an AI agent is heading off-track. Appian also automatically stops agents that exceed resource limits to protect system performance. For details, see Stop a running AI agent and Run limitations.

Unattended execution model

Process agents are designed to run autonomously without human intervention, completing multi-step workflows from start to finish. This unattended execution model enables true straight-through processing for complex tasks that previously required manual decision-making.

When you design an unattended process agent:

• Plan for autonomous operation: Structure your AI agent to complete its entire workflow without requiring human input during execution.
• Build in exception handling: Use your prompt to define when the AI agent should escalate to a human rather than proceeding. For example, "If the confidence score is below 70%, flag for manual review instead of proceeding."
• Design for side effects: Many AI agents don't return explicit outputs. Instead, they create side effects during execution such as updating records, sending notifications, or starting downstream processes.
• Enable human oversight: While AI agents run unattended, you can monitor their decisions and intervene when needed through the Monitor tab or by building review workflows.

This unattended model is what differentiates AI agents from AI skills, which are typically called within a process where you control the flow. AI agents control their own flow based on your instructions and the tools you provide.

Conversational execution model

Chat agents are designed to interact with end users through a live conversational interface. Rather than running autonomously in the background, a chat agent responds to user messages in real time, maintaining context across multiple exchanges within a session.

When you design a chat agent:

• Plan for multi-turn conversations: Structure your instructions so the chat agent can handle follow-up questions and build on prior context within a session.
• Design for user interaction: Chat agents wait for user input between responses. They guide users through tasks, answer questions, and provide recommendations conversationally.
• Use conversation persistence: End users can return to previous conversations and resume where they left off. For details, see Conversation Persistence for Chat Agents.

For guidance on embedding a chat agent in an interface, see Add a Chat Agent to an Interface.

For step-by-step guidance on creating AI agents, see Create and Configure an AI Agent. For process agent integration details, see Use AI Agents in a Process.

When to use an AI agent

AI agents are best suited for tasks that require flexible, goal-driven automation using dynamic reasoning. Unlike deterministic logic (where the same inputs always produce the same outputs), AI agents interpret context, adapt to variations, and make decisions that can't be fully scripted in advance. They use a prompt and tools to reason through various challenges, freeing you from having to model every possible scenario up front.

The strongest AI agent use cases share a common trait: the path to the solution is unpredictable. The AI agent must research, discover, and course-correct based on what it finds — rather than following a predetermined sequence of steps.

AI agents are the right choice when your use case meets one or more of these criteria:

• The next step depends on what the AI agent discovers. The AI agent searches across systems, interprets results, and decides where to look next. For example, an AI agent researching similar cases might find an initial set of candidates, then keep searching with different criteria until it's satisfied it found all the best matches — adapting its approach based on what each search reveals.
• The AI agent must adapt its approach based on what it finds. The AI agent pulls information from different systems, but unlike a process model that follows a fixed sequence of lookups, the AI agent decides which sources to check next based on what earlier steps revealed. For example, an AI agent might start by searching a knowledge base, realize the results are insufficient, and pivot to checking support tickets or policy documents — reasoning its way to a conclusion rather than following a predetermined path.
• The input is too large or complex for a single prompt. Simple unstructured inputs like short emails can be handled with a single AI skill. But when the input involves lengthy or complex documents — for example, five 300-page attachments — an AI agent can intelligently break down the content, extract relevant information across sections, and reason through what it finds to determine the next action.
• Dynamic routing depends on real-time context. The AI agent assigns tasks based on urgency, or case attributes without predefined routing conditions.
• Edge cases and exceptions need flexible handling. AI agents excel at pushing straight-through processing rates higher by handling the unpredictable variations that would require excessive branching in a process model.
• Orchestrating actions across external systems using MCP tools without building custom integrations.
• Specialized AI agents collaborate to complete complex workflows. A primary AI agent coordinates with separate AI agents for triage, data extraction, and approval routing.

For detailed examples and implementation guidance, see Use Cases. To learn how to build a complete triage solution, see Tutorial: Build a Triage AI Agent.

When not to use an AI agent

AI agents are powerful, but they're not the right solution for every automation challenge. Understanding when to use traditional process models or rules-based automation helps you build more reliable, cost-effective solutions.

Every AI agent run consumes AI actions for reasoning, planning, and tool selection — even when the path is predictable. If the path to the solution is known in advance, this reasoning adds cost and latency without adding value. Consider this cost-benefit trade-off when deciding whether an AI agent is the right approach.

Avoid using AI agents when:

• The process path is fully known and unchanging. If you can describe the AI agent's work as "Step 1, Step 2, Step 3" with no branching based on discovered context, a traditional process model is faster, more reliable, and less expensive to run. Even if you think your path is undefined or unpredictable, that's often because you haven't brought all stakeholders together to define the workflow. Once you do, you may realize the process is more deterministic than you originally thought.
• The AI agent follows the same path almost every time. If production data shows the AI agent takes the same route 99% of the time, convert that path to a deterministic process model. Reserve AI agents for the remaining edge cases.
• The task is a single cognitive operation. If the work involves one action — classify, extract, summarize, or generate — use an AI skill instead. AI agents are designed for multi-step reasoning, not single-task execution.
• High-volume, simple tasks need maximum efficiency. For repetitive tasks that don't require reasoning (such as data validation or simple routing), rules-based automation offers better performance and lower cost.
• Real-time performance is critical. AI agents involve LLM calls and tool execution, which introduce latency. For processes requiring sub-second response times, use traditional automation. Design AI agents to run upstream so results are ready when humans need them.

Understanding cost and reliability trade-offs

When choosing between AI agents, AI skills, and rules-based automation, consider both the execution characteristics and business requirements of your use case.

The cost of reasoning

AI agents reason before they act. Each run consumes AI actions for interpreting the prompt, evaluating available tools, planning the next step, and generating responses — even when the eventual path is straightforward. This reasoning cost is the trade-off you accept in exchange for flexibility and autonomy.

The reasoning cost is justified when the AI agent's path genuinely varies based on context. The reasoning cost is wasted when the AI agent follows a predictable sequence that could be encoded in a process model.

Reliability considerations

• Rules-based automation: Produces deterministic, 100% predictable results. The same inputs always generate the same outputs.
• AI skills: Highly reliable for their specific task, but outputs may vary slightly based on the LLM's interpretation.
• AI agents: Reliability depends on prompt quality, tool configuration, and the complexity of the reasoning required. More complex AI agents may need iterative refinement to achieve consistent results.

Cost considerations

AI agents consume resources differently than traditional automation:

• AI actions usage: Each LLM call consumes AI actions based on prompt length, tool descriptions, and response generation.
• Tool execution: Every tool call (querying records, starting processes, reading documents) adds processing time and resource consumption.
• Multi-step reasoning: Complex AI agents that require multiple reasoning cycles consume more resources than simple, linear processes.

To reduce unnecessary AI action consumption, pass data the AI agent always needs as inputs rather than requiring tool calls to retrieve it. Tool calls consume AI actions for reasoning about when and how to use them.

Development and maintenance

• Rules-based automation: Requires skilled developers to build and maintain, but produces highly reliable results.
• AI skills: Moderate development effort focused on prompt engineering and testing.
• AI agents: Lower barrier to entry for basic AI agents, but complex AI agents require careful prompt design, tool configuration, and ongoing monitoring to maintain performance.

Choosing the right automation approach

Appian offers multiple automation capabilities, each suited to different types of tasks. The key question when choosing an approach is: Is the path to the solution unpredictable?

• If the path is known and repeatable, use rules-based automation or a process model with AI skills.
• If the path varies based on what the AI agent discovers during execution, use an AI agent.
• If you're unsure, start with a process model. You can always introduce an AI agent for the reasoning-intensive portions of the workflow.

Decision framework

Approach	Best for	Reliability	Cost	Example use cases
Rules-based automation	Deterministic, high-volume tasks where accuracy is paramount	Very high (100% predictable)	Very low execution cost	Payroll calculations, financial transactions, compliance checks, data validation
AI skill	Single, well-defined cognitive tasks that benefit from AI	High for specific task	Low to medium	Document summarization, sentiment analysis, text classification, entity extraction
AI agent	Complex, multi-step processes requiring reasoning and dynamic decision-making	Medium to high (probabilistic)	Medium (AI actions + tool calls)	Case triage, lead qualification, multi-system research and synthesis, exception handling

When to use rules-based automation

Choose traditional process models and rules when:

• You need 100% deterministic results.
• The process logic is fully defined and stable.
• High-volume execution requires maximum efficiency.
• Strict audit trails are required.
• Real-time performance is critical.

When to use an AI skill

Choose AI skills when:

• You need AI to perform a single, repeatable task.
• The task involves text analysis, generation, or transformation.
• You control the process flow and specify exactly when AI is used.
• You want to add AI capabilities to an existing process.

When to use an AI agent

Choose AI agents when:

• The path to completion varies based on what the AI agent discovers.
• The AI agent must adapt its approach based on intermediate findings, rather than following a fixed sequence of steps.
• The logic requires multi-step reasoning that can't be handled in a single prompt or defined with static rules.
• Dynamic decision-making is required at multiple points in the workflow.
• Edge cases and exceptions require flexible, context-aware handling.

Hybrid approach

The most effective automation strategies combine multiple approaches:

• Use AI agents for research, reasoning, and decision-making where the path is unpredictable.
• Use AI skills for specific cognitive tasks (classify, extract, summarize) within a structured process.
• Use rules-based automation for deterministic execution and high-volume processing.
• Design for human oversight on exceptions and edge cases.

Deciding when to use an AI skill or an AI agent

This comparison helps you choose between AI skills and AI agents for specific scenarios.

Option	When to use it
AI skill	Choose this option when you need AI to perform a single, well-defined task within a structured process. You control the flow, define each step, and specify exactly when AI should be used—for example, to extract data, classify text, or summarize a document. AI skills support models in customer provided AWS accounts.
AI agent	Choose this option when you need AI to handle a complex, multi-step task where the process can't be fully defined up front. AI agents use your prompt and available tools to determine how to reach a goal, making them ideal for tasks that involve iterative tool use, adaptive decision-making across multiple steps, or coordination with other AI agents.

The following diagram shows how to decide whether to use an AI skill or an AI agent for your use case.

Key terms

Term	Description
AI agent	A configurable design object that performs tasks using a prompt and tools you define.
Process agent	An AI agent type that automates routine tasks without interruption as part of a business workflow. Runs via the Execute AI Agent smart service in a process model.
Chat agent	An AI agent type that engages users through natural conversation as a virtual assistant. Runs via the agent chat component in an interface.
Prompt	An instruction that tells the AI agent what task to perform and how to approach it. When you create a new AI agent, enter your prompt in the Instructions field.
Tool	A design object that allows the AI agent to interact with people, data, and systems or provides context for the task.
AI Agent tool	A tool that allows one AI agent to call another, enabling multi-agent collaboration.
MCP tool	An external tool provided by a Model Context Protocol (MCP) server that an AI agent can call at runtime. MCP tools let agents interact with third-party systems without custom integrations. See MCP Tools for AI Agents.
Input	Data the AI agent receives to complete a task. You configure inputs when you set up the AI agent.
Output	Data the AI agent returns after completing a task. You configure outputs when you set up the AI agent.

AI Guardrails and AI agents

AI Guardrails apply automatically to all AI agent interactions. Administrators configure environment-wide content safety policies in the Admin Console, and Appian enforces them on every AI agent execution without requiring changes to individual agent configurations.

When a Block action triggers during an AI agent conversation, the agent receives the administrator-configured block message instead of the AI model's response. The agent then proceeds with different content. The agent does not stop or fail—it treats the block message as input and continues its workflow.

Feedback